Skip to main content

Findings & Evidence

Better Harness is deliberately honest: every claim in a report is bounded by what was actually observed. This page explains the boundaries that make a report trustworthy.

The evidence boundary​

Three evidence domains stay independent until unified analysis:

  • Session evidence — what agents actually did in relevant Task Episodes, where the host supports a real local session source.
  • Project evidence — repository mechanisms: instructions, validation, hooks, delivery gates, and reusable knowledge routes.
  • Agent assets — configured Skills, commands, rules, MCP tools, and instruction files.

The shared boundary across all three: configured assets can establish that a mechanism exists, but only linked task evidence can establish that it was used or improved an outcome.

What a finding contains​

Each finding turns one supported gap into an actionable row:

  • Impact — why the gap matters for the work loop;
  • Evidence — the visible source the finding is bounded by;
  • Expected output — what "fixed" looks like;
  • Scoped repair — a bounded fix plan draft for review;
  • Acceptance checks — how to validate the repair landed.

Missing or partial evidence stays explicit in the report instead of becoming an unsupported score or claim.

The project lens: five software capabilities​

Alongside the task-centered Agent Work Loop, the independent project-evidence pass uses a static lens:

CapabilityQuestion it answers
Context MapCan the agent find the right context, boundary, risk area, and next step?
Environment ReadinessCan the project set up, run, reset, and diagnose without guesswork?
Fast FeedbackDo relevant checks return useful feedback quickly after a change?
Quality GatesAre architecture, security, schema, migration, and drift rules enforced?
Change SafetyAre risky actions, acceptance, and recovery controlled?

Unobserved task behaviour is never treated as missing repository capability.

Confidence​

Reports rate confidence Low/Medium/High, bound to how much was actually executed versus only read. Static-only first passes stay Low/Medium; passing a current check proves the intervention was exercised, and only a comparable later result can prove the loop improved.

Going deeper​